Over the past years, several semiconductor producers have lost customers due to yield problems. While producers tend to be hesitant to disclose specific yield figures, industry estimates commonly place these yields between 85% and 95%. Yield losses are particularly pertinent for new product generations, where recipes and production parameters must be calibrated from scratch. Consequently, determining the root causes of yield losses is not only a primary concern but also a potential competitive edge in the semiconductor industry.

Semiconductor manufacturing has long leaned on statistical methods for quality improvement. Many existing approaches are based on linear methods, such as the correlation analysis between equipment routing features and yield. However, these linear methods often fail to detect nonlinear relationships or interaction effects between process parameters. Overlooking these complexities, especially in manufacturing settings with large data quantities, can easily lead to missing the root causes of yield losses.

To address the shortcomings of linear methods, many producers have recently turned to data analysis based on conventional machine learning (decision trees, random forests, etc.). But there is an inherent limitation with these methods: they are designed for predicting quality outcomes rather than improving them. When assessing feature importance, machine learning algorithms can overfit the underlying data, which leads to misleading conclusions. They might assign high importance to variables that are simply correlated to yield issues, as opposed to being causally related. At the same time, critical process parameters influencing yield might be neglected, as these methods do not account for the sequential and hierarchical nature of the manufacturing processes.

Facing these challenges, our case company encountered these very limitations, struggling to pinpoint the root causes of its quality losses using both linear methods and conventional machine learning algorithms.